Modelpublic

skillsafeai/dreamshaper-8-lcm-onnx-webgpu

sourceHugging Facecreativeml-openrail-mupdated 3d agoView on Hugging Face

0likes

Model Card

DreamShaper 8 LCM, fp16 ONNX for the browser

Exported for onnxruntime-web over WebGPU from Lykon/dreamshaper-8-lcm at commit 4645d8bc6a8e6b106d21606d63e8460cdad4f1a6. No weights were retrained or modified - only exported.

File	Size	Graph
`text_encoder.onnx`	246 MB	CLIP ViT-L/14 text model, fp16, `input_ids` int64 [1,77] -> `last_hidden_state` float32 [1,77,768]
`unet.onnx`	1.72 GB	SD 1.5 UNet, fp16 inside, float32 IO: `sample` [1,4,64,64], `timestep` [1], `encoder_hidden_states` [1,77,768] -> `out_sample`
`vae_decoder.onnx`	198 MB	SD VAE decoder, fp32, divides by 0.18215 itself: `latent` [1,4,64,64] -> `image` [1,3,512,512] in [-1,1]

Standard ai.onnx operators, opset 17, no external data, single files (each under 2 GiB).

Sampling (LCM, 4 steps, no guidance)

timesteps = [999, 759, 499, 259]; a_t = alphas_cumprod (scaled_linear 0.00085..0.012)
x = randn(1,4,64,64)
for each t:  eps = unet(x, t, text_encoder(tokens))
             x0  = (x - sqrt(1-a_t) eps) / sqrt(a_t)
             den = c_out(t) x0 + c_skip(t) x        # sigma_data 0.5, timestep_scaling 10
             x   = last ? den : sqrt(a_next) den + sqrt(1-a_next) randn
image = vae_decoder(x) / 2 + 0.5

Checked against diffusers StableDiffusionPipeline + LCMScheduler (fp32, 4 steps, guidance 1.0) with the same random stream: PSNR 60.7 / 51.0 / 54.1 dB. In Chrome (WebGPU, Apple silicon): ~2.2-3.1 s per 512x512 image after the session build.

Licence and use restrictions

CreativeML Open RAIL-M, as the source model. Its use-based restrictions apply to this export and its outputs - including no content that exploits or harms minors, no non-consensual imagery of real people, no defamation, harassment or illegal use.